|Abstract:||Many languages disfavor coda voiced stops, but the number of ways in which languages resolve coda voiced stops is limited: i.e. languages alter voiced stops by devoicing but not by any other phonological means. For example, underlying /ab/ can become [ap], but not *[am], *[aba] or *[a]. To explain this observation, Steriade (2001/2008) claims that (i) speakers maximize the perceptual similarity between inputs and outputs, assuming that (ii) devoicing universally yields an outcome that is perceptually most similar to the original form. This paper reports a series of similarity judgment experiments to test the premise (the clause (ii)). The results are mixed: several orthography-based judgment studies of English speakers demonstrate that devoicing yields an outcome that is most similar to the target forms, compared to nasalization, deletion, or epenthesis. However, the auditory follow-up experiments reveal a more complex picture. The experiments overall provide support for the P-map hypothesis, but they also suggest a nuanced picture of what the underlying knowledge of similarity must be.